Notes on what we’re building, what broke, and what the testing actually showed. Written by the person doing the work.
Most AI products handle dangerous capabilities with permissions and prompt instructions. We took the capability out of the toolset entirely. Here’s what that means architecturally, and why “the model was told not to” isn’t a safety guarantee.
The model refused to invent numbers — and freely invented mechanisms. It called a deep order book “extremely thin” and gave real execution advice on it. The cause was one line of code discarding depth before the model ever saw it. What that taught us about where AI failures actually live.
When a feed has few trades, the easy message is “data may be delayed.” It’s also wrong, and it teaches traders to distrust the wrong thing. How we rewrote every data-quality message to say what’s actually true.
The difference isn’t the model — it’s whether the system holds your context by default, can act within visible limits, and notices things while you’re asleep.
Want these when they land? Apply for beta testing below — that’s where they go out first.
Beta testers get first access — and founding-member pricing when we launch. No card required. See pricing ›
We’ll only email you about beta access and launch. Unsubscribe anytime.